CN108985316B - Capsule network image classification and identification method for improving reconstruction network - Google Patents

Capsule network image classification and identification method for improving reconstruction network Download PDF

Info

Publication number
CN108985316B
CN108985316B CN201810509412.6A CN201810509412A CN108985316B CN 108985316 B CN108985316 B CN 108985316B CN 201810509412 A CN201810509412 A CN 201810509412A CN 108985316 B CN108985316 B CN 108985316B
Authority
CN
China
Prior art keywords
network
layer
vector
image
loss
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201810509412.6A
Other languages
Chinese (zh)
Other versions
CN108985316A (en
Inventor
段书凯
张金
邹显丽
王丽丹
耿阳阳
陆春燕
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Southwest University
Original Assignee
Southwest University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Southwest University filed Critical Southwest University
Priority to CN201810509412.6A priority Critical patent/CN108985316B/en
Publication of CN108985316A publication Critical patent/CN108985316A/en
Application granted granted Critical
Publication of CN108985316B publication Critical patent/CN108985316B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a capsule network image classification and identification method for improving a reconstruction network, which comprises the following steps: s1, constructing a capsule network; s2, inputting an image training set to the capsule network, and finishing image classification, identification and calibration after training and learning by the capsule network; s3, inputting the image to be classified to the capsule network, and outputting the vector v of the working networkjThe largest one of the median values is the obtained recognition result; s4, the capsule network outputs the recognition result of the image to be classified; wherein the reconstructed network structure of the capsule network is a deconvolution operation. Has the advantages that: a new reconstruction network structure is provided, the vector is restored into an image through deconvolution operation, network parameters are adjusted by comparing errors of the restored image and an original image, the number of calculation parameters is reduced, and more operation memories are left for hardware equipment.

Description

Capsule network image classification and identification method for improving reconstruction network
Technical Field
The invention relates to application of a capsule network in image classification, in particular to a capsule network image classification and identification method for improving a reconstruction network.
Background
In recent years, convolutional neural networks have been developed rapidly in the directions of image recognition, target detection, semantic segmentation, and the like, and generally comprise convolutional layers, active layers, pooling layers and full-link layers, where a pooling layer is an important component in a convolutional neural network, typically maximum pooling and average pooling operations, and can reduce the size of an input feature map and reduce the amount of computation of a model, but the pooling layer also has a problem of position information loss.
Aiming at the problem that position information is lost in a pooling layer in a convolutional neural network, Hinton in 2017 proposes a capsule network (capsnet), the capsule network adopts vectors as input and output, adopts a dynamic routing mechanism to update parameters, can extract position information, can extract more accurate characteristic information relative to the convolutional neural network, and is expected to replace the convolutional neural network structure at the present stage.
However, the existing capsule network design has the defect of large parameter quantity in the aspect of image processing, the model occupies a large memory, and the data volume processed by running hardware at the same time is small.
Disclosure of Invention
In order to solve the problem of large quantity of parameters of the capsule network, the invention provides a new reconstruction network structure for the existing capsule network, the vector is restored into an image through deconvolution operation, and the error of the restored image and the original image is compared to adjust the network parameters, so that the capsule network image classification and identification method for improving the reconstruction network is provided, the quantity of the calculated parameters is reduced, and more operation memories are left for hardware equipment.
In order to achieve the purpose, the invention adopts the following specific technical scheme:
a capsule network image classification and identification method for improving a reconstruction network comprises the following steps:
s1, constructing a capsule network, wherein the capsule network is provided with a working network and a proofreading network, the working network is used for inputting an image and outputting an identification result of the image, and the proofreading network is used for training and adjusting parameters of the working network;
the working network comprises a convolution structure and a full-connection structure, wherein the convolution output end of the convolution structure is connected with the full-connection input end of the full-connection structure, the convolution structure is a convolution layer and a PrimaryCaps layer which are sequentially connected, and the full-connection structure is a network structure which sequentially performs weight calculation, dynamic route adjustment and activation function operation;
the proofreading network comprises a margin Loss operation structure and a reconstruction network structure which are parallel, wherein a Loss input end of the margin Loss operation structure is connected with a full-connection output end of the full-connection structure, a reconstruction input end of the reconstruction network structure is respectively connected with the full-connection output end of the full-connection structure and a vector layer of an input image, a Loss output end of the margin Loss operation structure and a reconstruction output end of the reconstruction network structure are respectively connected with a Loss function input end of a Loss layer, and a Loss function output end of the Loss layer is connected with an optimization function calculation layer;
the reconstruction network structure comprises a Reshape layer, a deconvolution structure, a Flatten layer and a variance calculation layer which are sequentially connected, wherein the variance input end of the variance calculation layer is respectively connected with the Flatten layer and a vector layer of an input image, and the variance output end of the variance calculation layer is connected with the Loss function input end of a Loss layer;
s2, inputting an image training set to the capsule network, and finishing image classification, identification and calibration after training and learning by the capsule network;
s3, inputting the image to be classified to the capsule network, and outputting the vector v of the working networkjThe largest one of the median values is the obtained recognition result;
s4, the capsule network outputs the recognition result of the image to be classified.
The reconstruction network structure of the existing capsule network is a function operation of full connection of a plurality of layers, namely only vector transformation operation, and the operation amount is large, through the design, vector parameters are converted into image parameters for operation through a deconvolution structure, the parameter amount is reduced, but the performance such as actually obtained image processing accuracy is unchanged, and the like, so that the operation hardware of the capsule network can have larger memory margin.
Further, the specific process of the capsule network training and learning of step S2 is as follows:
s2.1, sequentially inputting the images in the image training set into a working network, and obtaining an output vector v after calculation of the working networkj；
S2.2, selecting the output vector v with the maximum parameter valuejInputting a margin loss operation structure and calculating to obtain a deviation value;
s2.3, the output vector v with the maximum parameter valuejInputting a reconstructed network structure and converting the reconstructed network structure into a feature map through a Reshape layer;
s2.4, performing deconvolution operation on the feature map by using a deconvolution structure to obtain a reconstructed image;
s2.5, converting the reconstructed image into a reconstructed vector through a Flatten layer;
s2.6, calculating the reconstructed vector and the input image vector through a sum variance calculation layer to obtain a variance vector;
s2.7, inputting the variance vector and the deviation value obtained in the step S2.2 into a Loss layer to obtain the Loss of the working network;
s2.8, the loss amount is optimized by an optimization function calculation layer and then fed back to a working network;
and S2.9, adjusting parameters of each layer from back to front in a reverse order by the working network until the identification accuracy of the working network is constant, and finishing the training and learning of the capsule network.
Output vector v of working networkjThe method comprises the steps that a plurality of vectors are provided, wherein the largest one of the vectors is an image classification recognition result of a working network, the recognition result has a certain error when the recognition result is not trained, training and learning are to reduce the error towards the correct direction, and finally an accurate recognition result is obtained.
Described further, the input of the margin loss operation structure is the output vector v of the working networkjThe output is sigmajLj，LjThe calculation formula of (a) is as follows:
Lj＝Tjmax(0,m+-‖vj‖)2+λ(1-Tj)max(0,‖vj‖-m-)2
wherein, TjFor the actual class of input image, m+Is | | | vjUpper boundary of | m-Is | | | vjThe lower boundary of | l, λ is the adjustment coefficient;
the Loss function of the Loss layer is calculated as the output sigma of the margin Loss operation structurejLjAnd adding the loss amount to the reconstruction error of the reconstructed network structure to obtain the loss amount.
The actual type of the input image and the recognition result are calculated to obtain the deviation amount in the above manner.
Further described, the deconvolution structure of the reconstructed network structure is 1 deconvolution layer, 1 convolution layer, and 2 deconvolution layers connected in sequence.
Through the alternative operation of convolution and deconvolution, the calculation distortion is reduced, and the error of the deconvolution structure is prevented from being large and the working network parameters are prevented from being adjusted wrongly.
Further describing, the deconvolution layers all adopt deconvolution operations with convolution kernels of 4 × 4 and step size of 2;
the convolutional layer adopts convolution operation with convolution kernel 2 x 2 and step length 1.
Described further, the weight of the full-link structure is calculated as:
Figure BDA0001671792050000051
the dynamic routing adjustment is:
Figure BDA0001671792050000052
Figure BDA0001671792050000053
Figure BDA0001671792050000054
wherein u isiInput vector, v, being a fully connected structurejIn order to output the vector, the vector is,
Figure BDA0001671792050000055
is a weight vector, WijAs weight parameter, bijFor dynamic routing parameters, cijFor the tuning parameters, k is the number of dynamic routing parameters, sjThe adjusted intermediate vector is dynamically routed.
Through the design, the full-connection structure also has dynamic routing adjustment capacity, namely the output vector vjDynamic adjustment to dynamic routing parameters bijThereby adjusting the calculation error.
Described further, the function of the activation function operation is:
Figure BDA0001671792050000056
wherein v isjAs an output vector, sjThe adjusted intermediate vector is dynamically routed.
The activation function is a new activation function, and the activation function is added into the calculation of the capsule network, so that the image classification and identification accuracy can be greatly improved.
Described further, the function of the activation function operation is:
Figure BDA0001671792050000057
wherein v isjAs an output vector, sjThe adjusted intermediate vector is dynamically routed.
The optimization function of the optimization function calculation layer is an Adam function.
The invention has the beneficial effects that: the new reconstruction network structure is provided, the vector is restored into an image through deconvolution operation, network parameters are adjusted by comparing the error of the restored image with the error of the original image, the number of calculation parameters is reduced, and more operation memories are left for hardware equipment; a new activation function is provided, so that the image classification and identification accuracy is greatly improved.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
FIG. 2 is a schematic diagram of the frame structure of a capsule network;
FIG. 3 is a schematic diagram of the training learning steps of the capsule network;
FIG. 4 is a schematic diagram of a capsule network structure according to the first embodiment;
fig. 5 is a schematic diagram of a reconfigurable network structure according to the first embodiment;
FIG. 6 is a diagram of training and testing analysis according to the first embodiment;
FIG. 7 is a schematic diagram of a conventional capsule network reconstruction network architecture;
FIG. 8 is a diagram of a training and testing analysis of a conventional capsule network;
FIG. 9 is a graph comparing the test results of the first embodiment with those of the conventional capsule network;
FIG. 10 is a diagram of capsule network training and test analysis after replacement of a new activation function;
FIG. 11 is a comparison graph of the test effect of the capsule network before and after replacement of the activation function.
Detailed Description
The invention is described in further detail below with reference to the following figures and specific embodiments:
as shown in fig. 1, a capsule network image classification and identification method for improving a reconstruction network comprises the following steps:
s1, constructing a capsule network;
s2, inputting an image training set to the capsule network, and finishing image classification, identification and calibration after training and learning by the capsule network;
s3, inputting the image to be classified to the capsule network, and outputting the vector v of the working networkjThe largest one of the median values is the obtained recognition result;
s4, the capsule network outputs the recognition result of the image to be classified.
The capsule network is provided with a working network and a proofreading network as shown in fig. 2, wherein the working network is used for inputting an image and outputting an identification result of the image, and the proofreading network is used for training and adjusting parameters of the working network;
the working network comprises a convolution structure and a full-connection structure, wherein the convolution output end of the convolution structure is connected with the full-connection input end of the full-connection structure, the convolution structure is a convolution layer and a PrimaryCaps layer which are sequentially connected, and the full-connection structure is a network structure which sequentially performs weight calculation, dynamic route adjustment and activation function operation;
the proofreading network comprises a margin Loss operation structure and a reconstruction network structure which are parallel, wherein a Loss input end of the margin Loss operation structure is connected with a full-connection output end of the full-connection structure, a reconstruction input end of the reconstruction network structure is respectively connected with the full-connection output end of the full-connection structure and a vector layer of an input image, a Loss output end of the margin Loss operation structure and a reconstruction output end of the reconstruction network structure are respectively connected with a Loss function input end of a Loss layer, and a Loss function output end of the Loss layer is connected with an optimization function calculation layer;
preferably, the reconstructed network structure in this embodiment includes a Reshape layer, a deconvolution structure, a scatter layer, and a variance calculation (SSE) layer, which are connected in sequence, where variance input ends of the sum variance calculation layer are respectively connected to the scatter layer and a vector layer of an input image, and a variance output end of the sum variance calculation layer is connected to a Loss function input end of a Loss layer, as shown in fig. 4 and 5;
as shown in fig. 4, the weight calculation of the full-connection structure is preferably:
Figure BDA0001671792050000071
the dynamic routing adjustment is:
Figure BDA0001671792050000081
Figure BDA0001671792050000082
Figure BDA0001671792050000083
wherein u isiInput vector, v, being a fully connected structurejIn order to output the vector, the vector is,
Figure BDA0001671792050000084
is a weight vector, WijAs weight parameter, bijFor dynamic routing parameters, cijFor the tuning parameters, k is the number of dynamic routing parameters, sjThe adjusted intermediate vector is dynamically routed.
In this embodiment, the function of activating function operation is:
Figure BDA0001671792050000085
wherein v isjAs an output vector, sjThe adjusted intermediate vector is dynamically routed.
Preferably, the input of the margin loss operation structure is the output vector v of the working networkjThe output is sigmajLj，LjThe calculation formula of (a) is as follows:
Lj＝Tjmax(0,m+-‖vj‖)2+λ(1-Tj)max(0,‖vj‖-m-)2
wherein, TjFor the actual class of input image, m+Is | | | vjUpper boundary of | m-Is | | | vjThe lower boundary of | l, λ is the adjustment coefficient;
the Loss function of the Loss layer is calculated as the output sigma of the margin Loss operation structurejLjAnd adding the loss amount to the reconstruction error of the reconstructed network structure to obtain the loss amount.
Preferably, the deconvolution structure of the reconstructed network structure in this embodiment is 1 deconvolution layer, 1 convolution layer, and 2 deconvolution layers connected in sequence;
preferably, the deconvolution layers all adopt deconvolution operations with convolution kernels of 4 × 4 and step length of 2;
the convolutional layer adopts convolution operation with convolution kernel 2 x 2 and step length 1.
The optimization function of the optimization function calculation layer is preferably an Adam function.
As shown in fig. 3, the specific process of capsule network training and learning in step S2 is as follows:
s2.1, sequentially inputting the images in the image training set into a working network, and obtaining an output vector v after calculation of the working networkj；
S2.2, selecting the output vector v with the maximum parameter valuejInputting a margin loss operation structure and calculating to obtain a deviation value;
s2.3, the output vector v with the maximum parameter valuejInputting a reconstructed network structure and converting the reconstructed network structure into a feature map through a Reshape layer;
s2.4, performing deconvolution operation on the feature map by using a deconvolution structure to obtain a reconstructed image;
s2.5, converting the reconstructed image into a reconstructed vector through a Flatten layer;
s2.6, calculating the reconstructed vector and the input image vector through a sum variance calculation layer to obtain a variance vector;
s2.7, inputting the variance vector and the deviation value obtained in the step S2.2 into a Loss layer to obtain the Loss of the working network;
s2.8, the loss amount is optimized by an optimization function calculation layer and then fed back to a working network;
and S2.9, adjusting parameters of each layer from back to front in a reverse order by the working network until the identification accuracy of the working network is constant, and finishing the training and learning of the capsule network.
The images identified in this embodiment are lung slices, wherein 2526 images contain malignant lung nodules, and 3967 images do not contain lung nodules and lung nodules that are benign, i.e., 6691 total images.
70% of the data set was used as the training set and 30% as the test set. Training set: the lung nodules are not included, the lung nodules are benign 2927, and the lung nodules are malignant 1756; and (3) test set: lung nodules were not included and were 1238 benign, 770 malignant lung nodules.
Fig. 6 is a diagram of analysis of training and testing data using the reconstructed network structure of the present invention, fig. 7 is a diagram of a reconstructed network structure of a conventional capsule network, that is, a reconstructed network structure using full connection operation, fig. 8 is a diagram of analysis of training and testing data thereof, and fig. 9 is a diagram of comparison of testing effects after training of both, which shows that the difference between both is very small.
But the parameter quantity of the two is smaller through the following comparison:
the reconstruction network structure of the invention:
inputting: 2 vectors of length 16;
reshape: a length 16 vector reshape becomes a 4 × 4 feature map: no parameters;
the 4 × 4 feature maps are subjected to deconvolution operation with a convolution kernel of (4,4) and a step size of 2, to obtain 64 feature maps with a size of 8 × 8: parameters 1 × 4 × 64 are 1024;
the 64 8 by 8 signatures are convolved (convolution kernel 2 by 2) to 64 7 by 7 signatures: the parameters 64 × 2 × 64 is 16384;
the 64 7 by 7 signatures were deconvoluted (convolution kernel 4 by 4) to 32 14 by 14 signatures: parameters 64 × 4 × 32 — 32768;
the 32 14 by 14 signatures were deconvoluted (convolution kernel 4 by 4) to 1 28 by 28 signatures: the parameters 32 × 4 × 1 are 512;
1 28 × 28 signature plots (compressed) into vectors of length 784: no parameters;
therefore, the total number of parameters of the new reconstructed network is: 1024+16384+32768+512 are 50688.
The reconstruction network structure of the full connection operation is as follows:
inputting: 2 vectors of length 16;
full connection 1: 1 vector of length 16 to vector of length 512: parameter 16 × 512 — 8192;
full connection 2: length 512 vectors to length 1024 vectors: parameter 512 x 1024 ═ 524288;
full connection 3: length 1024 vectors to length 784 vectors: parameter 1024 × 784 — 802816;
therefore, the total parameters of the original reconstructed network are: 8192+524288+802816 is 1335296.
On the basis of the above scheme, the second embodiment further designs a new activation function, that is, the function of the activation function operation is:
Figure BDA0001671792050000111
wherein v isjAs an output vector, sjThe adjusted intermediate vector is dynamically routed.
The data analysis of the capsule network training and testing after the new activation function is replaced is shown in fig. 10, and the comparison of the testing effects before and after the activation function is replaced is shown in fig. 11, which obviously shows that the recognition accuracy of the new activation function on the capsule network is greatly improved.

Claims (9)

1. A capsule network image classification and identification method for improving a reconstruction network is characterized by comprising the following steps:
s1, constructing a capsule network, wherein the capsule network is provided with a working network and a proofreading network, the working network is used for inputting an image and outputting an identification result of the image, and the proofreading network is used for training and adjusting parameters of the working network;
the working network comprises a convolution structure and a full-connection structure, wherein the convolution output end of the convolution structure is connected with the full-connection input end of the full-connection structure, the convolution structure is a convolution layer and a PrimaryCaps layer which are sequentially connected, and the full-connection structure is a network structure which sequentially performs weight calculation, dynamic route adjustment and activation function operation;
the proofreading network comprises a margin Loss operation structure and a reconstruction network structure which are parallel, wherein a Loss input end of the margin Loss operation structure is connected with a full-connection output end of the full-connection structure, a reconstruction input end of the reconstruction network structure is respectively connected with the full-connection output end of the full-connection structure and a vector layer of an input image, a Loss output end of the margin Loss operation structure and a reconstruction output end of the reconstruction network structure are respectively connected with a Loss function input end of a Loss layer, and a Loss function output end of the Loss layer is connected with an optimization function calculation layer;
the reconstruction network structure comprises a Reshape layer, a deconvolution structure, a Flatten layer and a variance calculation layer which are sequentially connected, wherein the variance input end of the variance calculation layer is respectively connected with the Flatten layer and a vector layer of an input image, and the variance output end of the variance calculation layer is connected with the Loss function input end of a Loss layer;
s2, inputting an image training set to the capsule network, and finishing image classification, identification and calibration after training and learning by the capsule network;
s3, inputting the image to be classified to the capsule network, and outputting the vector v of the working networkjThe largest one of the median values is the obtained recognition result;
s4, the capsule network outputs the recognition result of the image to be classified.
2. The capsule network image classification and identification method of the improved reconstruction network according to claim 1, wherein: the specific process of capsule network training and learning of step S2 is as follows:
s2.1, sequentially inputting the images in the image training set into a working network, and obtaining an output vector v after calculation of the working networkj；
S2.2, selecting the output vector v with the maximum parameter valuejInputting a margin loss operation structure and calculating to obtain a deviation value;
s2.3, the output vector v with the maximum parameter valuejInputting a reconstructed network structure and converting the reconstructed network structure into a feature map through a Reshape layer;
s2.4, performing deconvolution operation on the feature map by using a deconvolution structure to obtain a reconstructed image;
s2.5, converting the reconstructed image into a reconstructed vector through a Flatten layer;
s2.6, calculating the reconstructed vector and the input image vector through a sum variance calculation layer to obtain a variance vector;
s2.7, inputting the variance vector and the deviation value obtained in the step S2.2 into a Loss layer to obtain the Loss of the working network;
s2.8, the loss amount is optimized by an optimization function calculation layer and then fed back to a working network;
and S2.9, adjusting parameters of each layer from back to front in a reverse order by the working network until the identification accuracy of the working network is constant, and finishing the training and learning of the capsule network.
3. The capsule network image classification and identification method of the improved reconstruction network according to claim 1 or 2, characterized in that: the input of the margin loss operation structure is an output vector v of the working networkjThe output is sigmajLj，LjThe calculation formula of (a) is as follows:
Lj＝Tjmax(0,m+-‖vj‖)2+λ(1-Tj)max(0,‖vj‖-m-)2
wherein, TjFor the actual class of input image, m+Is | | | vjUpper boundary of | m-Is | | | vjThe lower boundary of | l, λ is the adjustment coefficient;
the Loss function of the Loss layer is calculated as the output sigma of the margin Loss operation structurejLjAnd adding the loss amount to the reconstruction error of the reconstructed network structure to obtain the loss amount.
4. The capsule network image classification and identification method of the improved reconstruction network according to claim 1, wherein: the deconvolution structure of the reconstruction network structure comprises 1 deconvolution layer, 1 convolution layer and 2 deconvolution layers which are connected in sequence.
5. The capsule network image classification and identification method of the improved reconstruction network according to claim 4, wherein: the deconvolution layers all adopt deconvolution operations with convolution kernels of 4 multiplied by 4 and step length of 2;
the convolutional layer adopts convolution operation with convolution kernel 2 x 2 and step length 1.
6. The capsule network image classification and identification method of the improved reconstruction network according to claim 1, wherein: the weight of the full connection structure is calculated as:
Figure FDA0001671792040000031
the dynamic routing adjustment is:
Figure FDA0001671792040000032
Figure FDA0001671792040000033
Figure FDA0001671792040000034
wherein u isiInput vector, v, being a fully connected structurejIn order to output the vector, the vector is,
Figure FDA0001671792040000035
is a weight vector, WijAs weight parameter, bijFor dynamic routing parameters, cijFor the tuning parameters, k is the number of dynamic routing parameters, sjThe adjusted intermediate vector is dynamically routed.
7. The capsule network image classification and identification method of the improved reconstruction network according to claim 1, wherein the function of the activation function operation is:
Figure FDA0001671792040000041
wherein v isjAs an output vector, sjThe adjusted intermediate vector is dynamically routed.
8. The capsule network image classification and identification method of the improved reconstruction network according to claim 1, wherein the function of the activation function operation is:
Figure FDA0001671792040000042
wherein v isjAs an output vector, sjThe adjusted intermediate vector is dynamically routed.
9. The capsule network image classification and identification method of the improved reconstruction network according to claim 1, wherein: the optimization function of the optimization function calculation layer is an Adam function.
CN201810509412.6A 2018-05-24 2018-05-24 Capsule network image classification and identification method for improving reconstruction network Active CN108985316B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810509412.6A CN108985316B (en) 2018-05-24 2018-05-24 Capsule network image classification and identification method for improving reconstruction network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810509412.6A CN108985316B (en) 2018-05-24 2018-05-24 Capsule network image classification and identification method for improving reconstruction network

Publications (2)

Publication Number Publication Date
CN108985316A CN108985316A (en) 2018-12-11
CN108985316B true CN108985316B (en) 2022-03-01

Family

ID=64542630

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810509412.6A Active CN108985316B (en) 2018-05-24 2018-05-24 Capsule network image classification and identification method for improving reconstruction network

Country Status (1)

Country Link
CN (1) CN108985316B (en)

Families Citing this family (31)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109727197B (en) * 2019-01-03 2023-03-14 云南大学 Medical image super-resolution reconstruction method
CN109801305B (en) * 2019-01-17 2021-04-06 西安电子科技大学 SAR image change detection method based on deep capsule network
CN109840560B (en) * 2019-01-25 2023-07-04 西安电子科技大学 Image classification method based on clustering in capsule network
CN110032925B (en) * 2019-02-22 2022-05-17 吴斌 Gesture image segmentation and recognition method based on improved capsule network and algorithm
CN110059730A (en) * 2019-03-27 2019-07-26 天津大学 A kind of thyroid nodule ultrasound image classification method based on capsule network
CN110059741B (en) * 2019-04-15 2022-12-02 西安电子科技大学 Image recognition method based on semantic capsule fusion network
CN110009097B (en) * 2019-04-17 2023-04-07 电子科技大学 Capsule residual error neural network and image classification method of capsule residual error neural network
CN110163489B (en) * 2019-04-28 2023-07-14 湖南师范大学 Method for evaluating rehabilitation exercise effect
CN110110668B (en) * 2019-05-08 2022-05-17 湘潭大学 Gait recognition method based on feedback weight convolutional neural network and capsule neural network
CN110084320A (en) * 2019-05-08 2019-08-02 广东工业大学 Thyroid papillary carcinoma Ultrasound Image Recognition Method, device, system and medium
CN110414317B (en) * 2019-06-12 2021-10-08 四川大学 Full-automatic leukocyte classification counting method based on capsule network
CN110399899B (en) * 2019-06-21 2021-05-04 武汉大学 Cervical OCT image classification method based on capsule network
CN110288555B (en) * 2019-07-02 2022-08-02 桂林电子科技大学 Low-illumination enhancement method based on improved capsule network
CN110502970A (en) * 2019-07-03 2019-11-26 平安科技（深圳）有限公司 Cell image identification method, system, computer equipment and readable storage medium storing program for executing
CN110309811A (en) * 2019-07-10 2019-10-08 哈尔滨理工大学 A kind of hyperspectral image classification method based on capsule network
CN112308089A (en) * 2019-07-29 2021-02-02 西南科技大学 Attention mechanism-based capsule network multi-feature extraction method
CN110458852B (en) * 2019-08-13 2022-10-21 四川大学 Lung tissue segmentation method, device and equipment based on capsule network and storage medium
CN110599457B (en) * 2019-08-14 2022-12-16 广东工业大学 Citrus huanglongbing classification method based on BD capsule network
CN111046916A (en) * 2019-11-20 2020-04-21 上海电机学院 Motor fault diagnosis method and system based on void convolution capsule network
CN111241958B (en) * 2020-01-06 2022-07-22 电子科技大学 Video image identification method based on residual error-capsule network
CN111291712B (en) * 2020-02-25 2023-03-24 河南理工大学 Forest fire recognition method and device based on interpolation CN and capsule network
CN111292322B (en) * 2020-03-19 2024-03-01 中国科学院深圳先进技术研究院 Medical image processing method, device, equipment and storage medium
CN111612030A (en) * 2020-03-30 2020-09-01 华电电力科学研究院有限公司 Wind turbine generator blade surface fault identification and classification method based on deep learning
CN111460818B (en) * 2020-03-31 2023-06-30 中国测绘科学研究院 Webpage text classification method based on enhanced capsule network and storage medium
CN111461063B (en) * 2020-04-24 2022-05-17 武汉大学 Behavior identification method based on graph convolution and capsule neural network
CN111626361B (en) * 2020-05-28 2023-08-11 辽宁大学 Bearing sub-health identification method for improving capsule network optimization hierarchical convolution
CN111931882B (en) * 2020-07-20 2023-07-21 五邑大学 Automatic goods checkout method, system and storage medium
CN112364920B (en) * 2020-11-12 2023-05-23 西安电子科技大学 Thyroid cancer pathological image classification method based on deep learning
CN112528165A (en) * 2020-12-16 2021-03-19 中国计量大学 Session social recommendation method based on dynamic routing graph network
CN113870241A (en) * 2021-10-12 2021-12-31 北京信息科技大学 Tablet defect identification method and device based on capsule neural network
CN114338093B (en) * 2021-12-09 2023-10-20 上海大学 Method for transmitting multi-channel secret information through capsule network

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107301640A (en) * 2017-06-19 2017-10-27 太原理工大学 A kind of method that target detection based on convolutional neural networks realizes small pulmonary nodules detection
CN107527318A (en) * 2017-07-17 2017-12-29 复旦大学 A kind of hair style replacing options based on generation confrontation type network model

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2007053648A2 (en) * 2005-10-31 2007-05-10 The Regents Of The University Of Michigan Compositions and methods for treating and diagnosing cancer

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107301640A (en) * 2017-06-19 2017-10-27 太原理工大学 A kind of method that target detection based on convolutional neural networks realizes small pulmonary nodules detection
CN107527318A (en) * 2017-07-17 2017-12-29 复旦大学 A kind of hair style replacing options based on generation confrontation type network model

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Capsules for Object Segmentation;Rodney LaLonde 等;《arXiv:1804.04241v1 [stat.ML]》;20180411;第1-9页 *
Fast Convergent Capsule Network with Applications in MNIST;Xianli Zou 等;《ISNN 2018: Advances in Neural Networks – ISNN 2018》;20180326;第3-10页 *

Also Published As

Publication number Publication date
CN108985316A (en) 2018-12-11

Similar Documents

Publication Publication Date Title
CN108985316B (en) Capsule network image classification and identification method for improving reconstruction network
EP3410392B1 (en) Image processing system and medical information processing system
Zhang et al. Learning deep CNN denoiser prior for image restoration
CN108898577B (en) Benign and malignant pulmonary nodule identification device and method based on improved capsule network
CN110490082B (en) Road scene semantic segmentation method capable of effectively fusing neural network features
CN111489364B (en) Medical image segmentation method based on lightweight full convolution neural network
CN111161306B (en) Video target segmentation method based on motion attention
CN112365514A (en) Semantic segmentation method based on improved PSPNet
CN109190511B (en) Hyperspectral classification method based on local and structural constraint low-rank representation
CN111160229B (en) SSD network-based video target detection method and device
CN112270366B (en) Micro target detection method based on self-adaptive multi-feature fusion
CN111738270A (en) Model generation method, device, equipment and readable storage medium
CN114299358A (en) Image quality evaluation method and device, electronic equipment and machine-readable storage medium
CN108765287B (en) Image super-resolution method based on non-local mean value
CN108921017B (en) Face detection method and system
CN114119669A (en) Image matching target tracking method and system based on Shuffle attention
CN116503671B (en) Image classification method based on residual network compression of effective rank tensor approximation
CN113724344A (en) Hyperspectral-based remote sensing image compression method
CN111666872A (en) Efficient behavior identification method under data imbalance
CN113838104B (en) Registration method based on multispectral and multimodal image consistency enhancement network
CN115578325A (en) Image anomaly detection method based on channel attention registration network
CN108734222B (en) Convolutional neural network image classification method based on correction network
CN111881994B (en) Identification processing method and apparatus, and non-transitory computer readable storage medium
Altınbilek et al. Identification of paddy rice diseases using deep convolutional neural networks
CN111914904A (en) Image classification method fusing DarkNet and Capsule eNet models

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant